AI & ML Quick Path

Checkpoints and readiness

A short quiz for each of the 7 parts. Use them to find gaps; they are self-checks, not a certificate.

Part 1 checkpoint: the big picture

  1. A company wants to sort support emails into 5 categories and has 20,000 past emails with their categories. Which kind of learning is this?

    Explanation

    There are past examples with the correct answer (the category), so it is supervised; the answer is a category, so it is classification.

  2. What should you build before any machine-learning model?

    Explanation

    A baseline tells you whether the model is worth its cost; without a measurement you cannot tell if anything improved.

  3. Where should an API key for a model provider live?

    Explanation

    Code and browser files are visible to others; secrets belong in configuration read at runtime.

Part 2 checkpoint: data

  1. You include "refund issued" as a feature for predicting whether a customer will complain. What is wrong?

    Explanation

    The feature is only known after the event you are predicting, so the model looks great in testing and fails in real use.

  2. A column "city" has 300 different values. What is a sensible first step?

    Explanation

    Models need numbers; grouping rare values keeps the number of columns manageable.

  3. Why look at the data (EDA) before training?

    Explanation

    Most model problems are data problems; exploring first finds them cheaply.

Part 3 checkpoint: machine learning

  1. A fraud model has 98% accuracy where 2% of transactions are fraud. Is it good?

    Explanation

    With rare classes, check recall and precision for the rare class and compare with a baseline.

  2. Training accuracy 1.00, validation accuracy 0.80. Most likely?

    Explanation

    A large gap between training and validation results means the model memorised the training data.

  3. For a table of customer data, which model family is usually the strongest first choice?

    Explanation

    For tabular business data, gradient boosting is usually accurate, fast and easy to explain; start there after a baseline.

  4. You are forecasting next month’s sales. How should you split the data?

    Explanation

    A random split lets the model see the future; a time split matches how the model will be used.

Part 4 checkpoint: deep learning

  1. You have 500 labelled photos of product defects. What is the best starting point?

    Explanation

    A pretrained model already knows general visual features, so a small dataset is enough to adapt it.

  2. Validation loss starts rising while training loss keeps falling. What do you do?

    Explanation

    Rising validation loss is overfitting; early stopping and regularisation address it.

  3. Which design is behind modern large language models?

    Explanation

    Transformers use attention so every token can use information from every other token, and they train efficiently on GPUs.

Part 5 checkpoint: language models

  1. What does a language model actually learn during pretraining?

    Explanation

    Next-token prediction on huge amounts of text is the training task; knowledge and skills emerge from it.

  2. Your company needs answers from its own frequently changing documents. Best first approach?

    Explanation

    RAG keeps knowledge up to date without retraining and lets answers cite sources; fine-tuning suits style and format.

  3. Why can a 7-billion-parameter model run on a laptop?

    Explanation

    At 4 bits, 7 billion weights need roughly 3.5 to 4 GB instead of about 14 GB at 16 bits.

  4. What is LoRA?

    Explanation

    LoRA trains only a small number of new weights, so fine-tuning fits on modest hardware.

Part 6 checkpoint: AI applications

  1. Where should permission checks happen in a RAG system?

    Explanation

    Text the user may not see must never enter the prompt; prompt rules can be bypassed.

  2. A model returns `Sure! {"priority": "high"}`. What should the application do?

    Explanation

    Model output is untrusted input; validate it before use.

  3. An agent can issue refunds. What is the safest design?

    Explanation

    Consequential actions need approval and least privilege enforced in code.

  4. You changed the prompt. How do you know the app did not get worse?

    Explanation

    A fixed evaluation set run on every change is how regressions are caught.

Part 7 checkpoint: production

  1. Why load the model once at service start-up?

    Explanation

    Loading a model can take seconds; doing it per request makes every call slow.

  2. The PSI for an important input is 0.5. What does it mean?

    Explanation

    Above about 0.25 indicates a significant shift; check the cause, then retrain or adjust.

  3. A new model scores slightly lower than production on the frozen test set. What should the promotion gate do?

    Explanation

    Only promote models that beat the champion by the agreed margin on the same test.

  4. The Python model service times out. What should the .NET application do?

    Explanation

    Every AI call needs a timeout, limited retries and a fallback so the business keeps running.

Interview readiness signals

Things you can demonstrate with your own capstones. Tick them to track your self-assessment (stored in this browser only).

AI engineer

ML engineer